Ponente
Descripción
Positron range (PR) is one of the most important sources of resolution degradation in Positron Emission Tomography (PET). Despite that, most PET reconstruction software do not provide a specific accurate PR correction (PRC) for radionuclides with large PR such as 68Ga (including only a PRC for 18F in water), which impacts the accuracy of the studies. We recently developed Deep-PRC, a fast and accurate PRC based on a Convolutional Neural Network (CNN), as a post-processing step for reconstructed PET images. In this work, we introduce Deep-PRC as a useful tool for the PET community.
Deep-PRC is based on the CNN U-NET architecture. It was trained with a large number of 68Ga and 18F PET/CT images obtained from realistic Monte Carlo (MC) simulations from an adapted version of penEasy. As PR is a local effect, 3D patches from the input volumes were used as training elements. The results obtained with the trained Deep-PRC were compared against a standard Richardson-Lucy (R-L) deconvolution algorithm using an isotropic gaussian kernel as a model of the PR blurring. Trained Deep-PRC models can be generated for each isotope and PET scanner and stored as HDF5 files. These models can be applied to correct PR effects in PET/CT studies using, as input, the reconstructed 3D PET and CT images in the most common medical imaging formats.
We present results for 68Ga simulated acquisitions for the preclinical Inveon PET/CT scanner. The initial 68Ga images and the PR-corrected ones obtained by Deep-PRC and the Richardson-Lucy method, were compared against the 18F ones (used as a reference). The Full-Width-At-Half-Maximum (FWHM) of a small hot region was used to evaluate the differences in the spatial resolution in each case, while the noise was evaluated on a uniform region. Qualitatively, there is a clear improvement in the image quality with the use of 3D patches for the training. Models for other radionuclides with large PR, such as 124I, are being developed, as well as an extensive application to preclinical and clinical studies.
Deep-PRC provides a fast and accurate PRC method to recover the resolution loss present in PET studies with radionuclides such as 68Ga that emit positrons with large PR. A stand-alone application can apply the pretrained models to reconstructed PET/CT images to improve their accuracy without compromising the image quality.
